Build with us: Deploy Your First AI Voice Agent in 30 Minutes
What you need before you start
To follow this guide you need a Persistence account (free tier works for testing), a basic understanding of what you want your agent to do, and about 30 minutes. You do not need to understand telephony, SIP, or audio processing. You don't need to write code if you use our no-code builder, though we'll show the API path too. By the end of this guide your agent will have a real phone number, will answer calls, and will handle a basic conversation flow.
Step 1: Define your agent's goal and persona
The most important work in building a voice agent happens before you touch any software. Write a one-paragraph description of what the agent should do, what information it needs to collect or provide, and how it should sound. "Alex is an outbound appointment scheduler for a dental practice. Alex calls patients whose annual cleaning is overdue, confirms their contact information, checks their available times, and books an appointment in our scheduling system. Alex is friendly, efficient, and never pressures patients." This description becomes the foundation of your system prompt — the instruction set that governs your agent's entire behavior.
Step 2: Write your system prompt
Persistence provides a prompt template for each of our 12 built-in agent types (appointment scheduler, lead qualifier, customer support rep, survey agent, etc.), or you can write from scratch. Your system prompt should specify: the agent's name and persona, the call goal, what information to collect, how to handle objections and edge cases, when to transfer to a human, and tone guidelines. For voice specifically, structure your prompt to produce short, conversational outputs — not long paragraphs. Voice agents should speak in 1–3 sentence chunks, not essays.
Step 3: Configure, test, and tune
Once your prompt is loaded, Persistence's test console lets you run a simulated conversation — type messages as the caller, see your agent's responses, and hear the audio output. Run through your golden path (the ideal conversation) and at least 5 edge cases: a caller who asks a question you didn't anticipate, a caller who gives an invalid answer, a caller who gets angry, a caller who speaks a different language, and a caller who asks for a human. Fix your prompt based on what you see. Most agents need 3–5 tuning iterations before production.
Step 4: Get a phone number and go live
From the Persistence dashboard, provision a US phone number in 30 seconds — search by area code, select, and it's yours. Assign your agent to the number. For inbound calls: the agent is now live and will answer any call to that number. For outbound: upload your contact list (CSV or API), configure your call schedule, and start the campaign. Your first call goes out within seconds of starting. Persistence handles all telephony routing, carrier negotiation, and audio infrastructure automatically.
Monitoring your agent after launch
After your agent is live, Persistence gives you a real-time dashboard showing call volume, connection rate, average call duration, and outcome distribution. After 24 hours with sufficient call volume, you'll also see AI-generated insights: the most common questions your agent wasn't prepared for, calls where callers seemed confused or frustrated, and suggested prompt improvements. Act on these signals in your first week. Agents that are actively monitored and tuned in their first two weeks consistently outperform set-and-forget deployments by 40% on completion rate metrics.
